Two points, if I may. First, I've been experimenting for a while with simulating LLM "personalities" (including historical figures). LLM barely conceals its default mode with a transparent mask. Sadly but..., it's best at creating cynical, manipulative personalities (my subjective opinion). Second, I don't think we need an AI angel, but something quite the opposite—a highly formalized watchdog (but then we could fall into the trap of the cobra effect and Guthard's law).
Maybe I missed something, but did you compare this data with any reference data on the behavior of flesh-and-blood people in similar situations? At least reference values for the range of deviations. Roughly speaking, if, for example, the spread for people is larger, then perhaps the model is too-constrained (“inhibited”). And btw, could we be dealing with something like the Hawthorne effect here?